X-Git-Url: https://iankelling.org/git/?a=blobdiff_plain;f=newns;h=cd2466b7b371469a8d1f7c478298ef0fea370b91;hb=b0239f31ed6ba299fbbde710c3ebb080bd5f8c12;hp=4fc887c6f038765d368d82286b763cd071b1cb0f;hpb=633e865aec9974d560bc7b6e3e526713ed8b41e9;p=newns

diff --git a/newns b/newns
index 4fc887c..cd2466b 100755
--- a/newns
+++ b/newns
@@ -14,23 +14,114 @@
 # limitations under the License.
 
 
-# Create a network namespace. Designed for use from systemd.
-
 [[ $EUID == 0 ]] || exec sudo -E "$BASH_SOURCE" "$@"
 
-cd "${BASH_SOURCE%/*}"
-source ../errhandle/errcatch-function
-source ../errhandle/bash-trace-function
-errcatch
+if [[ ! $ERRHANDLE_PATH ]]; then
+    ERRHANDLE_PATH=$(readlink -f "${BASH_SOURCE}")
+    ERRHANDLE_PATH=$(readlink -f ${ERRHANDLE_PATH%/*}/../errhandle)
+fi
+err_sourced=true
+for p in $ERRHANDLE_PATH/{errcatch-function,bash-trace-function}; do
+    if [[ -e $p ]]; then
+        source $p
+    else
+        err_sourced=false
+    fi
+done
+if $err_sourced; then
+    errcatch
+else
+    set -eE -o pipefail
+    trap 'echo "$0:$LINENO:error: \"$BASH_COMMAND\" returned $?" >&2' ERR
+fi
+
+usage() {
+    cat <<EOF
+usage: ${0##*/} [OPTS] start|stop NS_NAME
+Nat a network namespace. systemd friendly
 
+Also creates a mount namespace with a cloned /run/resolvconf.
 
-## begin arg parsing ##
-action=$1
-nn=$2 # network namespace / namespace name
-## end arg parsing ##
+-c, --create    Create a named network namespace. When running from
+                the same network namespace as pid 1, this is set automatically.
+                A systemd created private network is in a network namespace
+                different than pid 1.
+-n NETWORK      x.x.x /24 private network to use. If not specified, uses
+                the first one starting at 10.173.1
+-h, --help      Show this help and exit.
+
+From within a systemd network namespace, nat it to the outside. This
+would be called from ExecStartPre, and or subsequent units called with
+JoinsNamespaceOf= and PrivateNetwork=true.
+
+Also create a named mount namespace under /root/mount_namespaces, so we
+can alter some system config for this namespace. Subsequent systemd
+command lines would be prefixed with:
+
+/usr/bin/nsenter --mount=/root/mount_namespaces/NS_NAME
+
+Note, this means that they can't run as unpriveledged users, but once
+systemd 233 comes out, it will have a bind mount option from within unit
+files, so the mount namespace won't be needed for most use cases, and I
+will update the script to that the mount namespace not created unless a
+flag is passed in. Patch welcome to add that flag before then.
+
+A recommmended dependency of this script is my other repo named "errhandle",
+which prints stack trace on error, and calls a cleanup function:
+https://iankelling.org/git/?p=errhandle, set ERRHANDLE_PATH, or put it
+in a directory adjacent to the absolute, resolved directory this file is
+in.
+
+Background:
+
+If we aren't creating a named network namespace, to join the namespace
+with a shell, I use:
+nsenter -n -m -t \$(pgrep PROCESS_IN_NAMESPACE) bash
+
+Note: if I knew how to easily ask systemd what pid a unit has, i would
+do that.
+
+If we do create the netns, to join it with a shell, we can do
+/usr/bin/nsenter --mount=/root/mount_namespaces/NAME --net=/var/run/netns/NAME bash
 
-## begin sanity checking ##
+"ip netns new ..." also does a mount namespace, then bind
+mounts each file/dir in /etc/netns/NS_NAME to /etc/NS_NAME. Note,
+for openvpn having it's own resolv.conf by using it's user script which
+calls resolvconf, this doesn't help much. What we actually want to do is
+copy /run/resolvconf somehwere then bind mount it on top of
+/run/resolvconf.
 
+Note: for debugging, adding set -x is a pretty good option.
+
+Please email me if you have a patches, bugs, feedback, or republish this
+somewhere else: Ian Kelling <ian@iankelling.org>.
+EOF
+    exit ${1:-0}
+}
+
+
+#### begin arg parsing ####
+create=false
+temp=$(getopt -l help,create hcn: "$@") || usage 1
+eval set -- "$temp"
+while true; do
+    case $1 in
+        -c|--create) create=true; shift ;;
+        -n) network=$2; shift 2 ;;
+        -h|--help) usage ;;
+        --) shift; break ;;
+        *) echo "$0: Internal error!" ; exit 1 ;;
+    esac
+done
+if (( $# != 2 )); then
+    usage 1
+fi
+
+action=$1
+nn=$2 # namespace name
+#### end arg parsing ####
+
+#### begin sanity checking ####
 install_error=false
 if ! type -p ip &>/dev/null; then
     echo "please install the iproute2 package"
@@ -43,47 +134,75 @@ fi
 if $install_error; then
     exit 1
 fi
-
-##   end sanity checking ##
+####   end sanity checking ####
 
 
 v0=veth0-$nn
 v1=veth1-$nn
 ip_base=10.173
 
+if ! $create && [[ $(readlink /proc/self/ns/net) == "$(readlink /proc/1/ns/net)" ]]; then
+    create=true
+fi
 
+# make the default network namespace be named
 target=/run/netns/default
 if [[ ! -e $target && ! -L $target ]]; then
     mkdir -p /run/netns
-    # make the default network namespace be named
     ln -s /proc/1/ns/net $target
 fi
 
 
 ipd() { ip -n default "$@"; }
+if $create; then
+    # run ip in the network namespace
+    ipnn() { ip -n $nn "$@"; }
+else
+    # we are already in the network namespace and it's unnamed.
+    # run ip in the network namespace
+    ipnn() { ip "$@"; }
+fi
+# default network namespace exec
 dexec() { ip netns exec default "$@"; }
+# mount namespace exec
+mexec() { /usr/bin/nsenter --mount=/root/mount_namespaces/$nn "$@"; }
 
-# note: this script could be easily adapted to create a
-# netns instead of using the systemd created one.
-# ip netns add NAME
-# ip -n NAME link set dev lo up
 
-# random note, ip netns exec creates a mount namespace and
-# remaps /etc to /etc/netns/NAME.
+# background: head -n1 is defensive. Not sure if there is some weird feature
+# for 2 routes to be 0/0.
+gateway_if=$(ipd route list exact 0/0 | head -n1| sed -r 's/.*dev\s+(\S+).*/\1/')
+
+if [[ ! $gateway_if ]]; then
+    cat >&2 <<EOF
+$0: error: failed to find gateway interface. No output from:
+ipd route list exact 0/0 | head -n1| sed -r 's/.*dev\s+(\S+).*/\1/'
+output from "ipd route list exact 0/0":
+$(ipd route list exact 0/0)
+EOF
+    exit 1
+fi
 
-# head -n1 is defensive. I don't know if it's possible to have more
-# than one default route.
-gateway_if=$(ipd route list exact 0/0 | head -n1| sed -r 's/.*\s(\S+)\s*$/\1/')
 nat() { dexec iptables -t nat $1 POSTROUTING -o $gateway_if -j MASQUERADE \
               -m comment --comment "systemd network namespace nat"; }
 
+# d = default
+diptables-add() {
+    if ! dexec iptables -C "$@" &>/dev/null; then
+        dexec iptables -I "$@"
+    fi
+
+}
+
 find_network() {
+    if [[ $network ]]; then
+        return
+    fi
     found=false
     existing=false
     ips="$(ipd addr show | awk '$1 == "inet" {print $2}')"
-    for ((i=0; i <= 254; i++)); do
+    for ((i=1; i <= 254; i++)); do
         network=$ip_base.$i
-        if printf "%s\n" "$ips" | grep "^${network//./\\.}"; then
+        if printf "%s\n" "$ips" | grep "^${network//./\\.}" >/dev/null; then
             existing=true
         else
             found=true
@@ -93,38 +212,93 @@ find_network() {
 }
 
 start() {
+    find_network
+    if ! $found; then
+        echo "$0: error: no open network found"
+        exit 1
+    fi
 
+    #### begin mount namespace setup ####
     mkdir -p /root/mount_namespaces
     if ! mountpoint /root/mount_namespaces >/dev/null; then
         mount --bind /root/mount_namespaces /root/mount_namespaces
-        mount --make-private /root/mount_namespaces
     fi
+    # note: This is outside the mount condition because I've mysteriously
+    # had this become shared instead of private, perhaps it
+    # got remounted somehow and lost the setting.
+    mount --make-private /root/mount_namespaces
     if [[ ! -e /root/mount_namespaces/$nn ]]; then
         touch /root/mount_namespaces/$nn
     fi
     if ! mountpoint /root/mount_namespaces/$nn >/dev/null; then
-        unshare --mount=/root/mount_namespaces/$nn
+        # Here, we specify that we only want mount changes changes under
+        # this mountpoint to be propagated into the bind, but changes
+        # from within the bind do not propagate to outside the bind.
+        #
+        # slave is documented in.
+        # /usr/share/doc/linux-doc-4.9/Documentation/filesystems/sharedsubtree.txt.gz
+        # documentation on propagation is a bit weird because it
+        # confusingly talks about binds, namespaces, and mirrors (which
+        # seems to be just another name for bind), shared subtrees
+        # (which seems to a term for binds and namespaces), and does not
+        # properly specify whether the documentation applies to binds,
+        # namespaces, or both. Notably, propagation for binds is marked
+        # on the original mount point, and propagation for a mount
+        # namespace is marked on mounts within the namespace.
+        unshare --propagation slave --mount=/root/mount_namespaces/$nn /bin/true
     fi
 
+    ####   end mount namespace setup ####
 
 
-    find_network
-    if ! $found; then
-        echo "$0: error: no open network found"
-        exit 1
+    if $create; then
+        ip netns add $nn
+        ip -n $nn link set dev lo up
     fi
 
     echo 1 | dexec dd of=/proc/sys/net/ipv4/ip_forward 2>/dev/null
 
+    # docker helpfully changes the default FORWARD to drop...
+    diptables-add FORWARD -i $v0 -j ACCEPT
+    diptables-add FORWARD -o $v0 -m conntrack --ctstate RELATED,ESTABLISHED -j ACCEPT
+
+
     _errcatch_cleanup=stop
-    ip link add $v0 type veth peer name $v1
-    ip link set $v0 netns default
+    ipnn link add $v0 type veth peer name $v1
+    ipnn link set $v0 netns default
     ipd addr add $network.1/24 dev $v0
     ipd link set $v0 up
     nat -C &>/dev/null || nat -A
-    ip addr add $network.2/24 dev $v1
-    ip link set $v1 up
-    ip route add default via $network.1
+    ipnn addr add $network.2/24 dev $v1
+    ipnn link set $v1 up
+    ipnn route add default via $network.1
+
+    ###### begin setup resolvconf
+    resolv_copy=/root/resolvconf-$nn
+
+    # this condition should never happen, just coding defensively
+    if mexec mountpoint /run/resolvconf &>/dev/null; then
+        mexec umount /run/resolvconf
+    fi
+    cp -aT /run/resolvconf $resolv_copy
+    if ! mexec mount -o bind $resolv_copy /run/resolvconf; then
+        echo "error: resolv-conf bindmount failed"
+        exit 1
+    fi
+    # if running dnsmasq, we have 127.0.0.1 for dns, but it can't listen on the loopback
+    # in the network namespace, so adjust the address.
+    if mexec [ -s /run/resolvconf/interface/lo.dnsmasq ]; then
+        mexec sed --follow-symlinks -i "s/nameserver 127\..*/nameserver $network.1/" /run/resolvconf/interface/lo.dnsmasq
+        mexec resolvconf -u
+    fi
+    # background: if we did this in openvpn's resolv-conf script, we could guard it in
+    # if capsh --print|grep '\bcap_sys_admin\b' &>/dev/null
+    # and we could get $nn by
+    # config_basename=${config%%.*}
+    # config_basename=${config_basename##*/}
+    # but dnsmasq forces us to do it earlier.
+    ######  end setup resolvconf
+
 
 }
 
@@ -137,6 +311,19 @@ stop() {
     if ! $existing; then
         if nat -C &>/dev/null; then nat -D; fi
     fi
+    dexec iptables -D FORWARD -i $v0 -j ACCEPT ||:
+    if $create && [[ -e /var/run/netns/client ]]; then
+        ip netns del $nn
+    fi
+
+    # not sure this is necessary since we are tearing down the mount namespace
+    if mexec mountpoint /run/resolvconf &>/dev/null; then
+        mexec umount /run/resolvconf
+    fi
+
+    if mountpoint /root/mount_namespaces/$nn >/dev/null; then
+        umount /root/mount_namespaces/$nn
+    fi
 }
 
 case $action in